Meta releases its first frontier models: Muse Wonder and Muse Wonder Full

Meta Superintelligence Labs released two models on Monday, October 12, 2026: Muse Wonder and Muse Wonder Full. TechCrunch calls them Meta's first attempt at a frontier model, meaning a system that aims to compete with the best work from OpenAI, Google and Anthropic. The company itself says both models are at the state-of-the-art level. There is no independent confirmation yet.

What we know about the release

The initial announcement contains few facts. There are two models. They come from the superintelligence lab, not from the team that previously built Llama. They ship under a new name, Muse. For Meta this is a notable shift: for three years, the Llama brand was the company's main asset when it came to models.

The name Muse Wonder Full hints at the senior model in the lineup, and Muse Wonder is probably a lighter version. That is only a guess based on the names. In the material available to us, Meta does not disclose model sizes, architecture or the difference between the two versions.

• Release date: October 12, 2026
• Models: Muse Wonder and Muse Wonder Full
• Developer: Meta Superintelligence Labs
• Positioning: Meta's first frontier model, claimed SOTA level
• Not disclosed: parameter count, context length, API pricing, license, benchmark results

Why expectations for this release were high

Meta set up the lab in the summer of 2025, after the failure of Llama 4. The model came out in April of that year and drew criticism for the gap between its claimed and actual results. In response, Zuckerberg invested about $14.3 billion in Scale AI for a 49% stake and put its founder, Alexandr Wang, in charge of the effort. A hiring spree targeting researchers at rival labs followed.

More than a year has passed since then. Muse Wonder is the first public result of that overhaul. That is why the model will be judged more harshly than a typical release: the market wants to know whether the money and the hiring paid off.

What to check before believing the claim

The words "state-of-the-art" cost little in a press release. Four things will carry more weight.

First: whether the weights are open. Llama built its reputation on an open ecosystem, and the answer determines whether Meta stays attractive to developers who build products on local models. Second: independent measurements on real-world tasks, including coding and agentic scenarios, rather than only on benchmarks the company chose itself. Third: token pricing and inference stability under load. Fourth: how quickly Meta builds Muse into its own products, where it has billions of users.

What teams and companies should do now

Don't migrate production on the strength of a single announcement. A better approach is to assemble your own set of 30–50 real queries that you currently use to evaluate models, and run them through Muse Wonder and Muse Wonder Full as soon as access opens. Compare quality, latency and cost on your own data.

Teams that depend on Llama should find out how Meta plans to support the previous lineup, and keep a fallback option. Those choosing a vendor from scratch should wait for the license: it determines whether you can fine-tune the model and deploy it on your own servers.

The main question for this release is whether third-party tests will confirm the claimed level. So far, the claim exists only on paper.